Reproduce “SpikingGamma: Surrogate-Gradient Free and Temporally Precise Online Training of Spiking Neural Networks with Smoothed Delays” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Roel Koopman, Sebastian Otte, Sander Bohté
SpikingGamma trains feedforward spiking neural networks online using cascades of leaky buckets to retain differently delayed traces of past input. It sends training errors through continuous signal estimates rather than spike events, avoiding surrogate gradients and backpropagation through time. On the 20-class SHD task at 3,000 frames, it reaches 92.42 ± 0.31%, versus 81.36 ± 1.11% for a reproduced method; this comparison is not fully matched.
Key insights
- Learnable delay buckets concentrate temporal responses around task-relevant events, producing more precise delay tuning than surrogate-gradient BPTT in controlled experiments.
- The fixed number of filter states makes theoretical memory usage independent of sequence length, allowing the model to tolerate much finer temporal discretization than several online baselines.
- Normalization, rectification, and the initialization of the bucket response are important for stable learning; gain regularization can increase sparsity with a limited accuracy trade-off.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.